good robot

@typesafeai has officially cooked

congrats @hackgoofer @CompleteSkeptic @EGafni - this is really elegant

major narrative violation

@jon3k Yes. As I said it alters tail logprobs in a narrow distribution, thereby changing the output.

LLM classic right here

This is a good one. Don't agree with every methodology (e.g. LLMaJ). But the taxonomy and the focus on failure modes at the seams is the important bit

Timely reading for anyone building agent harnesses in domain specific / OOD problem spaces

from @DrScottClark

Has software engineering really changed that much though?

Snow depth by Year

Temperature by Year

GPU-accelerated plot engine written from scratch.

Visualize 5M+ points at 60 FPS on a standard MBP.

Announcing Probably: the verifiable data agent

We raised a $9M seed round led by @a16z and @Accel to build an AI agent that can ingest, transform, analyze and visualize raw data.

Single take, unedited video analyzing 1B+ row NOAA weather data from Clickhouse 👇

BREAKING: Fable 4 still on track for 2027 release after dodging US export controls due to off-by-one error in their favor

@vansickn @MattRogish here's a draft teaser. will publish when our blog goes live in general which should be any day now

This is actually such a big deal

Just hang in there guys

The slop rate tops out at the max output tok/s

You just need to review exponentially faster

wagmi

Under-discussed problem right now with most frontier coding models. A leading contributor to slop and incidental complexity and daily pain

Great read @vansickn !

@depotdev I see what you did there

@jhoangeth Exhibit B: Probably, the data agent (aka data harness), crushes it easily in about 60 seconds, for 1/100th of the token cost

@jhoangeth Exhibit A: Codex (running in Pi with some data skills (like DuckDB), attempts to answer a simple question over a CSV and fails after 5+ minutes)

Meta prompt of the year.

incremental force direction with repulsion that converges to a steady state

make no mistakes

distsys pop quiz

what database produces this telemetry pattern and what it is doing in that moment?

well yes that IS the top news story thank you x

me waiting to stop the clanker the moment it tries to slip some slop by me

@ClickHouseDB Unreported incident. Quietly back online now :/

@ClickHouseDB plase halp

@ClickHouseDB cloud having issues right now? No status updates but something feels a bit off...

This is only going to get worse. People are starting to wake up to the reality.

The reasons are straightforward, but the rabbit hole runs deep... 🧵

Hallucination-free data analysis at the speed of thought.

Dirty data. Messy schemas. Mixed types. Impossible joins. Doesn't matter.

Next 250 waitlist entries releasing tomorrow.

Anyone can operate with the data intelligence of an F1000 enterprise now. For pennies.

I bet Codex it was wrong about a root cause.

Please to report it honored the bet

@AKWilk It's a real problem. We fixed it though

@jessfraz codex would have been like

@IamMiaChase enjoy lol

Ever since I was a wee lad I wanted a data agent that natively understands and efficiently manipulates tabular file formats at scale

Using an AI that _actually_ understands and manipulates raw data feels like science fiction.

Weather analysis at Clickhouse scale, for pennies

The difference between Codex and Opus

@dwarkesh_sp @collision

Elite prompt hack to stop Opus 4.5 over-engineering 👇

We made this bet last year. If you would like to purchase the future, it is now available for sale.

We have solved some of the hardest problems in building a "data agent" outside any major AI lab. If you want to experience an AI that can drop into raw, messy data, at massive scale with no guidance whatsoever and just "figure it out". Hit me up. The visuals are stunning too.

If anyone is interested in buying one of these. Check out probably (dot) dev - works very well even at massive scales

swisssshhh

Singularity achieved internally.